speechbrain.lm.counting module

N-gram counting, discounting, interpolation, and backoff

Authors

  • Aku Rouhe 2020

Summary

Functions:

ngrams

Produce all Nth order N-grams from the sequence.

ngrams_for_evaluation

Produce each token with the appropriate context.

pad_ends

Pad sentence ends with start- and end-of-sentence tokens

Reference

speechbrain.lm.counting.pad_ends(sequence, pad_left=True, left_pad_symbol='<s>', right_pad_symbol='</s>')[source]

Pad sentence ends with start- and end-of-sentence tokens

In speech recognition, it is important to predict the end of sentence and use the start of sentence to condition predictions. Typically this is done by adding special tokens (usually <s> and </s>) at the ends of each sentence. The <s> token should not be predicted, so some special care needs to be taken for unigrams.

Parameters:

  • sequence (iterator) – The sequence (any iterable type) to pad.

  • pad_left (bool) – Whether to pad on the left side as well. True by default.

  • left_pad_symbol (any) – The token to use for left side padding. β€œ<s>” by default.

  • right_pad_symbol (any) – The token to use for right side padding. β€œ</s>” by default.

Returns:

A generator that yields the padded sequence.

Return type:

generator

Example

>>> for token in pad_ends(["Speech", "Brain"]):
...     print(token)
<s>
Speech
Brain
</s>
speechbrain.lm.counting.ngrams(sequence, n)[source]

Produce all Nth order N-grams from the sequence.

This will generally be used in an N-gram counting pipeline.

Parameters:

  • sequence (iterator) – The sequence from which to produce N-grams.

  • n (int) – The order of N-grams to produce

Yields:

tuple – Yields each ngram as a tuple.

Return type:

None

Example

>>> for ngram in ngrams("Brain", 3):
...     print(ngram)
('B', 'r', 'a')
('r', 'a', 'i')
('a', 'i', 'n')
speechbrain.lm.counting.ngrams_for_evaluation(sequence, max_n, predict_first=False)[source]

Produce each token with the appropriate context.

The function produces as large N-grams as possible, so growing from unigrams/bigrams to max_n.

E.G. when your model is a trigram model, you’ll still only have one token of context (the start of sentence) for the first token.

In general this is useful when evaluating an N-gram model.

Parameters:

  • sequence (iterator) – The sequence to produce tokens and context from.

  • max_n (int) – The maximum N-gram length to produce.

  • predict_first (bool) – To produce the first token in the sequence to predict (without context) or not. Essentially this should be False when the start of sentence symbol is the first in the sequence.

Yields:

  • Any – The token to predict

  • tuple – The context to predict conditional on.

Example

>>> for token, context in ngrams_for_evaluation("Brain", 3, True):
...     print(f"p( {token} |{' ' if context else ''}{' '.join(context)} )")
p( B | )
p( r | B )
p( a | B r )
p( i | r a )
p( n | a i )